Hallucinations in GPT-2 Trained Model
Bibliographic record
Abstract
This paper analysis the phenomenon of "hallucinations" in text generated by GPT-2 when it produces irrelevant or illogical content.This work has quantified the extent of those hallucinations and look into ways of their mitigation.By using two main techniques: cosine similarity and frequency analysis.These techniques calculate coherency and relevance in the text produced by OpenAI GPT-2 at different training levels.Where a study case was implemented to train the model and ask the questions and retrain the model using these replays.The main findings indicate that this model hallucinates much less at the beginning of learning, with the situation significantly improving as training progresses.Extreme learning does not eliminate all such inadequacies, and more over-training led to more hallucinations.The hallucinated items span from smaller deviations to major content-wise deviations.An inspection reveals some patterns and cues that are predictive of increased output unreliability of the model.This research suggests a stricter training program that involve varied data sets to reduce the rate of hallucinations.More importantly, improve the accuracy of the model by reaching superior levels through the embedding of contextual and factual anchoring systems as well as designing algorithms for higher trigger identification.Other recommendations of the paper include post-generation text evaluation and continuous research to enhance the complexity of the models.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.008 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.001 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.004 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".